科研之友
微信
新浪微博
Facebook
分享链接
摘要
An integrated online process monitoring, fault diagnosis,and quality prediction framework based on multiple models is developed for multiphase batch processes. Two types of different-level PLS models are designed for each phase. In the first level, multiple PLS models based on variable unfolding, termed MVPLS, are employed for online monitoring and revealing the dynamic progressing direction that the quality will take at each time. In each phase, MVPLS model is established by performing PLS on variablewise data rearrangement, which can focus on the process information more relevant to the final quality from an overall phase-specific point of view. Moreover, control limits of quality prediction are explored and, thus, employed in process monitoring. Combined with conventional Hotelling-T-2 and SPE statistics, they can enhance the process understanding and monitoring performance. Once an abnormality is detected, the contributions of process variables to quality prediction are calculated in combination with control limits to help check the fault variables quantitatively. In the second level, regression models are designed based on the phase-specific average score trajectory for more credible prediction of end-of-batch quality, where the phase-specific average trajectory reveals the accumulation effects of process variations on quality in a simpler structure. For a normal batch, whenever a phase is successfully completed, a more accurate quality prediction can be naturally obtained by implementing the second-level regression analysis. The case studies from a simulated fed-batch penicillin cultivation process demonstrate the power and advantage of the proposed method.
